The blog post introduces 'Diffusion Elites', a novel algorithm that leverages pre-trained diffusion models to optimize search processes in high-dimensional spaces. The method combines the strengths of evolutionary search technique...
About:
Christian S. Perone is the author of the website 'Terra Incognita'. The site focuses on topics such as machine learning, evolutionary computation, genetic algorithms, genetic programming, Python, Arduino, math, science, deep learning, machine learning engineering, large language models, and natural language processing.
Website:
Specializations:
Incoming Links:
Subscribe to RSS:
The post discusses the development of VectorVFS, a software designed to store data embeddings directly in the filesystem without requiring an external database. It highlights the use of extended attributes (xattr) in major Linux f...
This opinion article discusses Gilbert Simondon's 1958 work 'On the Mode of Existence of Technical Objects,' highlighting his philosophy of technology and its relevance to contemporary discussions on AI and machine learning. The a...
The blog post details the development of a prototype panel for the TorchStation, designed to monitor GPU usage and memory during distributed model training. The author discusses challenges faced, particularly with power requiremen...
The article discusses the connection between differential geometry and score-based generative models, focusing on the metric tensor and its connection with the Stein score. It explains the concept of manifolds, tangent spaces, and...
0Torch Titan distributed training code analysis
2024-08-21 •
The post is a detailed analysis of the distributed training code in Torch Titan. It covers various interesting aspects such as disabling and controlling Python's garbage collector, custom GPU memory monitoring, custom profiling co...
The text is an experiment to share a memory-mapped tensor from PyTorch to Numpy, Jax, and TensorFlow in CPU without copying, ensuring changes in memory by torch are reflected on all shared tensors. The author demonstrates the proc...
The text discusses the concept of schematism in Kant's critique and its relevance to machine learning. It explains the forms of thought, categories, and transcendental deduction in Kant's philosophy. It also explores the connectio...
O post aborda as enchentes devastadoras que atingiram o Rio Grande do Sul em 2024, com foco em compreender a escala e o impacto desses eventos usando fotos de satélite e dados recentes. O autor destaca a importância da coleta de d...
0PyTorch 2 Internals – Talk
2023-12-11 •
...
The text discusses the significance of curved geometries and the methods for measuring things that don’t reside on orthonormal bases. It explores the connection between Riemannian metrics and diffusion/score matching in the contex...
The article provides an overview of the data pipelines used to create datasets for training large language models (LLMs). It focuses on the pipeline described in CCNet and LLaMA's paper, both from Meta. The pipeline includes stage...
Feste is a free and open-source framework for composing NLP tasks using a graph execution model that optimizes and executes tasks with automatic parallelization and batching. It automatically fuses multiple calls to the same backe...
The text shares recent publications on uncertainty estimation for deep neural networks and autonomous driving planning. The first publication discusses a method called L2M that uses the gradient second moment to construct Laplace ...